Sherpa

Sherpa is a hybrid web agent: a planner reads the page (screenshot + compact controls/content), a screenshot-only grounder picks click coordinates, and Playwright acts. Default models are Qwen3.5-35B-A3B (planner/verifier) and UI-TARS-1.5-7B (grounder) via OpenRouter.

Install (Browser Use–style)

With uv (recommended):

# From GitHub (until published on PyPI):
uv tool install git+https://github.com/AnilPuram/Sherpa.git

# Or from a local checkout:
uv tool install .

# After PyPI publish:
# uv tool install sherpa-agent
# uvx --from sherpa-agent sherpa --help

# One-shot without a permanent install:
uvx --from git+https://github.com/AnilPuram/Sherpa.git sherpa --help

Then:

sherpa install          # Playwright Chromium (+ Linux system deps)
sherpa init             # creates .env if missing
# edit .env → OPENROUTER_API_KEY=
sherpa "Confirm the page heading on https://example.com, then finish."

sherpa install runs the same pattern as Browser Use (uvx playwright install chromium, with --with-deps on Linux). If Chromium is missing when you run a task, Sherpa installs it automatically.

Dev checkout

uv sync
uv run sherpa install
uv run sherpa init
uv run sherpa "Confirm the page heading on https://example.com"

Usage

Config is one required key. Sherpa loads .env automatically (without overriding variables already set in your shell). Optional model/budget overrides are commented in .env.example.

# Start URL can be in the task, or passed explicitly:
sherpa "Confirm the page heading" --url https://example.com

# Headless + machine-readable result:
sherpa "…" --url https://example.com --headless --json

By default the browser is headed, steps print as they run, and artifacts go to artifacts/runs/<timestamp>/.

Verify offline

uv run ruff check .
uv run pytest
uv run sherpa eval

Use --real-model only when you intend paid OpenRouter calls for grounding eval:

uv run sherpa eval --real-model --output artifacts/eval.json

Benchmarks / research

WebVoyager subset tooling, judgment rubrics, run history, and the cross-agent bakeoff live under eval/ and the docs below. Live WebVoyager still requires --real-model so offline scoring stays safe and free.

uv run sherpa webvoyager --output artifacts/webvoyager-offline.json
uv run sherpa webvoyager --real-model --max-cost-usd 1.00 \
  --artifacts artifacts/webvoyager --output artifacts/webvoyager-live.json

See WEBVOYAGER_TESTS_AND_RESULTS.md, WEBVOYAGER_RUN_HISTORY.md, and eval/bakeoff-round2-comparison.md.

How it works

The loop keeps a bounded progress ledger and memories, recovers from grounding/execution errors, and requires a separate verifier to accept every done answer. Defaults: 20 planner steps, 5 consecutive corrections, high planner reasoning effort. Details: ARCHITECTURE.md.

Files

• src/sherpa/cli.py: simple task CLI + research subcommands
• src/sherpa/install_browser.py: sherpa install / auto Chromium setup
• src/sherpa/agent.py: bounded progress, recovery, and hybrid-observation agent loop
• src/sherpa/models.py: planner, grounder, and verifier OpenRouter boundary
• src/sherpa/eval.py: point-in-box grounding evaluation
• src/sherpa/webvoyager.py: offline validation and bounded live WebVoyager subset runner
• ARCHITECTURE.md: agent flow and component diagrams
• eval/bakeoff-round2-comparison.md: Sherpa vs Browser Use vs Magnitude bakeoff
• scripts/bakeoff/: external-agent bakeoff runners